Computer Architecture · All levels
Accelerator Architectures Tricky Q&A
20+ senior Accelerator Architectures interview questions.
Q&A bank
Answer with mechanism, pitfall, validation check, and product judgment. At 10+ years, a correct definition is not enough.
Senior answer rubric
Start with the failing metric and analysis context.
Explain the microarchitectural or system mechanism.
Name the cheapest evidence-gathering experiment.
Choose a bounded fix and state what it can regress.
Close with product tradeoff, validation, or escalation criteria.
How do you pick an accelerator pattern for a new workload family?
diagram
[INT][ARCH]
Q: How do you pick an accelerator pattern for a new workload family?
A:
Start with kernel structure, reuse behavior, control divergence, and software stack constraints before ranking patterns.
FOLLOW-UP TRAP: Selecting pattern by peak throughput marketing numbers.What is a red flag during pattern feasibility review?
diagram
[INT][ARCH]
Q: What is a red flag during pattern feasibility review?
A:
When utilization claims assume ideal tensor shapes and ignore realistic sparsity or batching distributions.
FOLLOW-UP TRAP: Accepting synthetic best-case utilization as baseline.Why can dataflow choice dominate performance more than clock frequency?
diagram
[INT][ARCH]
Q: Why can dataflow choice dominate performance more than clock frequency?
A:
Because dataflow determines movement and synchronization overhead, which often caps effective throughput.
FOLLOW-UP TRAP: Treating dataflow as a minor implementation detail.When is hybrid dataflow worth the complexity?
diagram
[INT][ARCH]
Q: When is hybrid dataflow worth the complexity?
A:
When workload diversity causes one fixed mode to underperform across meaningful deployment scenarios.
FOLLOW-UP TRAP: Adding hybrid modes without measurable cross-workload benefit.How do you diagnose compute-idle periods in an accelerator trace?
diagram
[INT][ARCH]
Q: How do you diagnose compute-idle periods in an accelerator trace?
A:
Correlate idle windows with transfer stalls, queue starvation, and phase synchronization events.
FOLLOW-UP TRAP: Assuming idle periods imply weak compute microarchitecture.What metric best exposes locality quality for architecture decisions?
diagram
[INT][ARCH]
Q: What metric best exposes locality quality for architecture decisions?
A:
Bytes moved per useful operation across hierarchy levels, paired with energy impact.
FOLLOW-UP TRAP: Relying on cache hit rate alone.Why can improved SRAM hit rate still fail KPI targets?
diagram
[INT][ARCH]
Q: Why can improved SRAM hit rate still fail KPI targets?
A:
Because off-chip traffic, coherency overhead, or burst inefficiency can still dominate end-to-end cost.
FOLLOW-UP TRAP: Declaring locality solved once hit rate increases.What should a senior integration checklist include beyond protocol compliance?
diagram
[INT][ARCH]
Q: What should a senior integration checklist include beyond protocol compliance?
A:
QoS behavior, coherency latency, power-state transitions, telemetry, and error recovery contracts.
FOLLOW-UP TRAP: Checking only functional correctness in isolation.How do you evaluate programmability risk in architecture selection?
diagram
[INT][ARCH]
Q: How do you evaluate programmability risk in architecture selection?
A:
Estimate compiler/runtime complexity, fallback path quality, and debuggability before RTL commitment.
FOLLOW-UP TRAP: Deferring software concerns until silicon bring-up.When should an accelerator be non-coherent by design?
diagram
[INT][ARCH]
Q: When should an accelerator be non-coherent by design?
A:
When workload and software model can tolerate explicit synchronization with clear bandwidth/latency benefits.
FOLLOW-UP TRAP: Defaulting to coherence without traffic analysis.What is the first triage for platform-level accelerator jitter?
diagram
[INT][ARCH]
Q: What is the first triage for platform-level accelerator jitter?
A:
Audit arbitration, queue isolation, interrupt handling, and firmware scheduling under concurrency.
FOLLOW-UP TRAP: Tuning kernel code before checking system arbitration.How do you avoid overfitting architecture to one benchmark suite?
diagram
[INT][ARCH]
Q: How do you avoid overfitting architecture to one benchmark suite?
A:
Use representative workload families, mixed concurrency scenarios, and holdout cases in acceptance.
FOLLOW-UP TRAP: Optimizing solely for public benchmark rankings.Why should accelerator decisions include verification cost?
diagram
[INT][ARCH]
Q: Why should accelerator decisions include verification cost?
A:
Verification scope and corner-case complexity can dominate schedule and risk for advanced architectures.
FOLLOW-UP TRAP: Assuming verification effort scales linearly with block size.What indicates a dangerous dataflow-memory mismatch?
diagram
[INT][ARCH]
Q: What indicates a dangerous dataflow-memory mismatch?
A:
Frequent spill/refill churn, burst fragmentation, and rising backpressure despite high core utilization.
FOLLOW-UP TRAP: Interpreting high MAC utilization as universal success.How do you frame area-power-performance tradeoffs to leadership?
diagram
[INT][ARCH]
Q: How do you frame area-power-performance tradeoffs to leadership?
A:
Present bounded options with KPI deltas, risk profile, software impact, and schedule confidence.
FOLLOW-UP TRAP: Giving only one preferred design without alternatives and risks.What is the role of observability in accelerator integration?
diagram
[INT][ARCH]
Q: What is the role of observability in accelerator integration?
A:
It provides deterministic post-silicon diagnosis for queues, coherency, and policy behavior.
FOLLOW-UP TRAP: Treating telemetry as optional debug overhead.When do you reject an integration signoff despite green block KPIs?
diagram
[INT][ARCH]
Q: When do you reject an integration signoff despite green block KPIs?
A:
Reject when platform contracts fail under concurrency, power-state transitions, or security isolation checks.
FOLLOW-UP TRAP: Passing integration because standalone kernels are fast.How should memory hierarchy be co-designed with dataflow?
diagram
[INT][ARCH]
Q: How should memory hierarchy be co-designed with dataflow?
A:
Jointly optimize tile policy, residency, bank access, and transfer cadence for target workload phases.
FOLLOW-UP TRAP: Designing hierarchy and dataflow in separate silos.What distinguishes senior answers in accelerator architecture interviews?
diagram
[INT][ARCH]
Q: What distinguishes senior answers in accelerator architecture interviews?
A:
They tie mechanism to measurable KPI impact, regression surface, and release decision criteria.
FOLLOW-UP TRAP: Reciting architecture buzzwords without ownership logic.What should happen after a successful accelerator optimization lands?
diagram
[INT][ARCH]
Q: What should happen after a successful accelerator optimization lands?
A:
Re-profile full system workloads to identify the next dominant bottleneck and update guardrails.
FOLLOW-UP TRAP: Assuming one successful optimization closes the project.Q&A drill guide
diagram
MECHANISM → METRIC → EXPERIMENT → TRADEOFF → VALIDATIONSketch while answering
diagram
ACCELERATOR DATAFLOW
Host CPU ── commands ──► Queue / scheduler
▲ │
│ completion ▼
Coherent memory ◄── DMA ── Local SRAM ──► Compute array
▲ │
└ tiles ┘
Peak TOPS matters only when data reaches the array at the needed rate.Key takeaways
Connect every architecture claim to a workload and measurable metric.
State verification and PPA impact before proposing design changes.
Common pitfalls
Feature-driven design without MPKI/IPC/bandwidth evidence.
Ignoring coherency and NoC traffic in cache and accelerator sizing.